Skip to content
Back to skills

Hunt Bizlogic

ASecurity

Business-logic flaw hunting - workflow/state bypass, price/quantity tampering, negative/overflow values, coupon/refund abuse, mass assignment, and logic races. The top-paying bug class with no scanner coverage. Wiki-first, FIND schema output.

  • 323 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 19, 2026
ai-agentspythonrustgoawstesting

Works with

  • cli
  • mcp

Security analysis

A100/100

Scanned September 19, 2026

npx -y skills add Encod3d-Sec/TORCH --skill hunt-bizlogic --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Hunt Bizlogic?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Hunt Bizlogic
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/encod3d-sec-hunt-bizlogic/badge)](https://www.skillsdirectory.com/skills/encod3d-sec-hunt-bizlogic)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: hunt-bizlogic
description: Business-logic flaw hunting - workflow/state bypass, price/quantity tampering, negative/overflow values, coupon/refund abuse, mass assignment, and logic races. The top-paying bug class with no scanner coverage. Wiki-first, FIND schema output.
---

# Hunt: Business Logic

**Assumes `hunt-core`** for the scope gate, two-account rule, confirmation gate, enumeration limits, stop conditions, wiki protocol, FIND output, and Deadends. Do not re-derive any of that here.

## Wiki

```
qmd_query "business logic flaw workflow state bypass price quantity tampering coupon refund race condition" via wiki-search MCP
```

Hub: [[web-moc]] (live index). Primary page: [[business-logic]].
Anchors: [[race-conditions]].

## No scanner finds these

This is the top-paying bug class and the one with **zero scanner coverage** - it is all manual reasoning. Find an assumption the developer made and break it: that steps happen in order, that values are positive, that the client cannot change a field, that an action cannot be repeated. Map the intended workflow first (read the feature, the happy path, every state transition); the flaw is always the gap between that and what the server actually enforces.

## Attack surface - rank before testing

Money and state-machine endpoints pay; cosmetic ones do not. Work them in this order:

1. **Checkout / payment** - client-supplied price, total, tax, currency, shipping, anywhere the server should recompute a value and instead trusts it.
2. **Refund / coupon / gift card / store credit** - re-redemption, stacking, negative amounts, a refund exceeding the purchase.
3. **Multi-step workflows** - checkout, transfer, signup, KYC, password reset, subscription: skip, replay, or reorder a step.
4. **Quantity / price fields** - negative, zero, overflow, decimal and rounding.
5. **State transitions** - an action valid only in one state performed from another (ship before pay, use before verify).

## Methodology

**Setup:** two accounts per `hunt-core` where the flaw is cross-actor; one account suffices for pure value or state tampering. Drive the load-bearing requests through **Burp Repeater** for operator visibility.

1. **Map the flow.** Enumerate every step and state of the target feature. Note each parameter and which server check governs it. The flaw is a check the client can reach around.

2. **Step / state bypass.** Skip a step (POST straight to the final endpoint), replay an earlier step, run steps out of order, reuse a one-time token, reach step N without completing N-1.

3. **Value tampering.**
```
quantity = -1            # negative -> credit / refund
price / amount = 0.01    # client-supplied price
currency swap            # pay in a weaker currency, credited in a stronger one
integer overflow / very large qty
decimal / rounding (0.001 * 1000)
```

4. **Repetition / limits / logic race.** Apply a coupon twice, redeem a gift card twice, exceed a per-account limit. When the limit is a check-then-act with no lock, fire a **small concurrent burst** at the single endpoint (logic [[race-conditions]]) so several requests pass the check before any commits. Keep the burst bounded: a handful of parallel requests is the proof, honor `no_dos`, never a sustained flood. Per `hunt-core` this is a bounded active proof, not object enumeration, so the 5-to-20 identifier ceiling does not apply, but the RoE rate cap does. Send the request to **Turbo Intruder** via `send_to_intruder` / the Burp MCP and fire it with a single-packet or synchronized-gate script so the operator watches the race live; a hand-rolled parallel `curl` loop hides it.

5. **Mass assignment / parameter injection.** Add fields the UI never sends (`isAdmin`, `role`, `balance`, `verified`, `discount`, `userId`) to JSON or form bodies; observe a privilege or state change.

6. **Trust-boundary confusion.** Values the server should compute but trusts from the client (total, tax, role, KYC status, account tier); flags re-sent and re-trusted on a later request.

7. **Identity / authorization logic.** An action allowed for the wrong account state (an unverified user performing a verified-only action) or against another user's object (-> [[access-control]] / `hunt-idor`).

## When the value is rejected

A rejected tamper is not a closed door. Retry the same value in a representation the server parses but the validator missed: an alternate field the server also reads, parameter pollution (`price=10&price=0.01`), a nested wrapper (`{"item":{"price":0.01}}`), a type change (string vs number, array vs scalar), sign and precision tricks (`-0`, `1e-2`, trailing decimals), or re-sending a server-computed field an earlier step trusted. Reorder the steps so validation runs before the value is set.

## Chaining

Logic flaws compound. A price or quantity tamper that touches *another user's* cart is logic + IDOR (hand off to `hunt-idor`); a coupon or limit bypass that only works under concurrency is logic + race ([[race-conditions]]). Chain them for higher impact and say so in the title - "race-condition coupon re-redemption" reads bigger than "coupon bug".

## Confirmation gate

**NOT confirmation:** the app accepted an odd value; a `200` on a tampered request; the tampered field echoed back in the response; an order or total that merely *displays* wrong on your screen; a coupon that "applied" without the charge changing.

**IS confirmation:** the unintended end-state is realized and verified downstream - a negative or inflated balance that persists, a purchase completed at the free or discounted price (order confirmed, payment captured at the tampered amount), a workflow step skipped **and** its downstream effect confirmed (account provisioned, funds moved, access granted). For a race, the limit is actually exceeded (two redemptions both settled). Re-verify in a clean session, and confirm from the state only the successful abuse produces, not from the request the server merely accepted.

## Severity

Rated on realized impact per `hunt-core`.

| Outcome | Typical |
|---|---|
| Direct financial theft, auth bypass, unlimited privilege | critical |
| Discount / refund abuse, quota bypass with money impact | high |
| Limited abuse needing preconditions | medium |

## Distill

Confirmed and reusable (a generic logic-abuse pattern, no client specifics): `python3 scripts/wiki-stage.py --kind technique --slug <slug> --target-page techniques/web/business-logic.md`.

## Deadends

```
Append: - [ ] logic <feature> on <host> -- server recomputes price/total, enforces the state
              machine, ignores extra fields, idempotency keys present; race lost after bounded burst
```

Record what the server enforced, so the next pass does not retest a solid boundary.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.